“We should probably be doing video.”
Somebody says it in the marketing meeting, everyone nods, and then nothing happens for another quarter. Nobody disagrees, and nobody schedules anything.
The reason it stalls is usually the same. The blog got restructured for AI, the sources went in, the numbers got checked — but the channel has three-digit subscribers and videos nobody watches, so video feels like a thing you earn the right to do later.
Video, in my head, was a numbers game. An unwatched video is a stack of brochures left on a chair nobody sits in.
Then a hundred million AI citations got opened up, and a lot of the videos the models pulled from were exactly the unwatched ones.
What Kind Of YouTube Video Does AI Actually Cite?
Mostly the readable ones, not the popular ones. 40.83% of the YouTube videos cited in AI answers had under 1,000 views, and views, likes, and subscriber counts all correlated with citation frequency at within ±0.03 — statistically nothing. The one number that moved was description length, at 0.31.
So the crowd and the model end up on different videos, and the model goes with whatever it can read.
The data comes from a study OtterlyAI published on March 2, 2026 — a company that sells monitoring for how brands show up in AI search.
They took 30 days of observed AI citations (over 100 million of them), pulled out the ones pointing at YouTube, and counted what those videos had in common. Six assistants were in scope: ChatGPT, Google AI Overviews, Google AI Mode, Perplexity, Microsoft Copilot, and Gemini.
The Two Numbers That Didn’t Collapse To Zero
So here’s the popularity side of the ledger:
- Views: −0.03; likes: −0.02; subscribers: −0.03
- Video length and title length: 0.02 each
Those are — a number between −1 and +1 for how tightly two things move together. Near zero means “unrelated”; you need something around 0.3 before you’d call it even a weak link.
Two numbers cleared that bar. Description length came in at 0.31, and recency (how new the video is) at 0.3.
Every proxy for popularity flatlined, and what survived was machine-readable substance plus freshness. That ordering is the interesting part.
Chapters Work Like Headings, And Only Google Reads Them
Some shape facts first. 94% of the cited videos were long-form and only 5.7% were Shorts — but half of them ran under eight minutes, so “longer is better” isn’t the lesson either. Descriptions averaged 334 words; titles averaged 19.
Then the number I keep coming back to: timestamps.
Timestamps are the “00:00 Intro” lines you drop into the description box, which split the video into chapters.
Only 31% of cited videos had them. But of the videos that did, 78% got cited more than once.
And chapter-level citation — where the answer points at a single chapter inside the video — showed up only in Google’s products: 73% in AI Overviews, 27% in AI Mode, and zero observed anywhere else.
Think of a chapter list as the H2s of a web page. A long document with no headings is hard for a person to skim and hard for a machine to index, which is really the same problem twice.
So Which Assistants Does Video Work Actually Reach?
Where these citations land is lopsided enough to change how you budget:
- Perplexity: 38.7%
- Google AI Overviews: 36.6%, Google AI Mode: 19.6%
- ChatGPT: 4.4% (Microsoft Copilot 0.5%, Gemini 0.2%)
Google’s surfaces plus Perplexity account for about 95% of it. ChatGPT barely cites video at all.
So video work and ChatGPT work aren’t the same line in the budget, and I’d stop pretending otherwise.
For scale: social and video platforms together are only 5.54% of all AI citations. Inside that slice, Reddit takes 46.4% and YouTube 31.8% — second place among social sources.
What This Data Can’t Tell You
The caveats, honestly, before anyone takes 40.83% to a planning meeting:
- OtterlyAI sells AI-search visibility monitoring, so “measure your AI citations” is a conclusion with money attached
- It’s correlation, not causation — nobody lengthened a batch of descriptions and watched what happened
- The sample is videos that were already cited; there’s no control group of uncited ones, so this describes the composition of the winners, not the odds of winning
- One 30-day snapshot, with no separation of seasonality or algorithm updates
- No language breakdown was published, so treat it as mostly-English evidence
I first read this as “fine, pad the descriptions” — and it doesn’t support that. What it supports is narrower: the videos getting cited have thick descriptions and chapter structure.
Maybe that’s selection all the way down: the videos the models were going to pick up anyway might just be the ones whose descriptions someone bothered to write properly. But the shape is consistent enough that I’d act on it before waiting for the randomized version.
Fix The Description Before You Reshoot Anything
Video visibility in AI isn’t a popularity contest you wait to win. It’s a formatting job on assets you already own, and a channel with three-digit view counts is still eligible for the citation.
Two things, and only two. Put at least three chapters in the description of your important videos, starting at 00:00 and running in ascending time order. And write the description out properly — a real summary, the product and category names, the relevant links.
Neither one needs a camera.
Then watch video separately from your blog numbers. Since nearly all of this lands in Google’s surfaces and Perplexity, a report built around ChatGPT will show you a flat line while the thing is actually moving.
Next time you open your channel, scroll past the view count and read the box underneath.
